Siddhant Academy

Comprehensive Guide to Self-Organizing Map (SOM)

Siddhant Bhattarai's photo
Siddhant Bhattarai
·

4 min read

1. What is a Self-Organizing Map (SOM)?

A Self-Organizing Map (SOM) is an unsupervised neural network used for dimensionality reduction, clustering, and data visualization. It projects high-dimensional data onto a lower-dimensional grid (typically 2D), preserving the topological relationships of the original data.

  • Key Idea: SOM organizes similar data points closer together on the map, forming clusters.

  • Inspiration: SOM is inspired by the way neurons in the brain organize themselves spatially to represent incoming stimuli.

2. How SOM Works

Algorithm Steps:

  1. Initialization:

    • A grid of neurons is initialized, each associated with a weight vector of the same dimension as the input data.

    • These weights are initialized randomly.

  2. Input Selection:

    • A random input vector from the dataset is chosen.

  3. Best Matching Unit (BMU) Identification:

  4. Weight Update:

  5. Neighborhood Shrinking and Learning Rate Decay:

    • Over time, the size of the neighborhood around the BMU decreases.

    • The learning rate α(t) also decreases.

  6. Iteration:

    • Steps 2–5 are repeated for several iterations until convergence.

3. Mathematical Principles Behind SOM

Distance Measure:

The most common distance measure used to find the BMU is the Euclidean distance:

Learning Rate Decay:

Neighborhood Function:

4. Key Factors to Consider Before Using SOM

  1. Grid Size: The size of the SOM grid impacts the level of detail in clustering. Larger grids can capture more complex patterns.

  2. Initialization: Proper weight initialization improves convergence.

  3. Learning Rate and Neighborhood Size: These parameters must decay over time to ensure convergence.

  4. Number of Iterations: Ensure enough training iterations for the SOM to stabilize.

  5. Distance Metric: Choosing an appropriate distance metric (e.g., Euclidean) is crucial for meaningful clustering.

5. Types of Problems Solved by SOM

  • Clustering: Grouping similar data points based on their features.

  • Dimensionality Reduction: Reducing high-dimensional data to 2D or 3D for visualization.

  • Anomaly Detection: Identifying outliers or unusual data points.

  • Pattern Recognition: Finding patterns in unlabeled data.

6. Applications of SOM

  • Customer Segmentation: Grouping customers based on purchasing behavior.

  • Healthcare: Clustering patient profiles based on medical history.

  • Marketing: Identifying customer preferences for targeted advertising.

  • Biological Data Analysis: Grouping gene expression profiles.

  • Fraud Detection: Detecting unusual patterns in transaction data.

7. Advantages and Disadvantages of SOM

Advantages

  • Unsupervised Learning: Can cluster data without labeled outputs.

  • Data Visualization: Projects high-dimensional data onto 2D or 3D grids for easy interpretation.

  • Topology Preservation: Retains the topological structure of the data.

  • Versatility: Works with different types of datasets (numerical, categorical).

Disadvantages

  • Parameter Sensitivity: Requires careful tuning of grid size, learning rate, and neighborhood size.

  • Computationally Intensive: Training large SOMs on large datasets can be slow.

  • Fixed Grid Size: The grid size must be predefined, which may limit flexibility.

  • Limited Interpretability: While SOM provides clusters, interpreting what the clusters represent can be subjective.

8. Performance Metrics for SOM

  1. Quantization Error: Measures the average distance between input vectors and their corresponding BMUs:

  2. Topographic Error: Measures how well the SOM preserves the topology of the input space.

  3. Silhouette Score: Evaluates how well data points fit within their assigned clusters.

  4. Davies-Bouldin Index: Measures the compactness and separation of clusters.

9. Python Code Example: SOM for Customer Segmentation

Below is an example of using SOM to cluster customers based on purchasing behavior.

Python Code

import numpy as np
import matplotlib.pyplot as plt
from sklearn.datasets import make_blobs
from minisom import MiniSom

# Generate sample data
X, y = make_blobs(n_samples=300, centers=4, random_state=42)

# SOM Initialization
som_grid_size = (10, 10)  # 10x10 grid
input_dim = X.shape[1]  # Number of features
som = MiniSom(som_grid_size[0], som_grid_size[1], input_dim, sigma=1.0, learning_rate=0.5)

# Randomly initialize weights
som.random_weights_init(X)

# Train SOM
som.train_random(X, num_iteration=1000)

# Visualize the SOM
plt.figure(figsize=(10, 10))
plt.title('Self-Organizing Map (SOM) Visualization')
for i, x in enumerate(X):
    w = som.winner(x)
    plt.plot(w[0] + 0.5, w[1] + 0.5, 'bo')  # Plot the BMU
plt.show()

Explanation of the Code

  • Dataset: A synthetic dataset of customer data.

  • MiniSom Library: A lightweight Python library for SOM implementation.

  • Grid Size: 10x10 SOM grid.

  • Training: The SOM is trained for 1000 iterations.

Expected Output

  • A 2D grid showing clusters of similar data points.

  • Each blue dot represents a Best Matching Unit (BMU), indicating where similar input data points map on the SOM grid.

10. Summary

Self-Organizing Maps (SOM) are powerful tools for unsupervised learning, clustering, and dimensionality reduction. They excel at projecting high-dimensional data onto 2D or 3D grids, making data visualization intuitive and insightful. SOMs are widely used in applications such as customer segmentation, healthcare, and anomaly detection. However, they require careful tuning of parameters like learning rate, grid size, and neighborhood function for optimal performance.

By mastering SOM, you can effectively cluster and visualize complex datasets and discover hidden patterns.